A detector answers a narrower question
An AI-detection tool might look for a signed provenance record, a vendor watermark, a platform record or statistical patterns. Those tests do not ask the same question. “No metadata found” does not mean “human-made”; “classifier score is high” is not a signed account of how a file was created.
Even a valid Content Credential has a scope: it can establish that a particular signer made an intact claim about a file. It cannot by itself establish that the depicted event is true. The C2PA explainer makes that distinction.
Why results can disagree
A generated image may lose portable metadata when copied into a new file, while an embedded watermark remains detectable. Conversely, a file may carry provenance without the detector for a vendor-specific watermark finding anything. A content classifier can make errors because it infers from patterns rather than checking a signed record. NIST's synthetic-content report treats provenance, watermarking and content-based detection as different approaches with different limitations.
For text, the stakes are especially clear: OpenAI withdrew an AI-written-text classifier because of low accuracy. That historical example does not measure every current detector, but it shows why a score should not be used as a verdict about a person.
A better evidence trail
Keep the original file and the source context. Check any signed credential with a validator. Use a watermark detector only for the vendor it covers, and treat classifier scores as leads for investigation rather than proof. If a platform has a label, read its explanation separately. The comparison of signal types helps choose the right check.
Free AI Wiper reports supported portable traces and verifies selected removal. It does not assign an AI-authorship probability.